Validation of an automated algorithm to consistently divide bi-atrial geometries from multiple imaging modalities according to the 15-segments bi-atrial model

Introduction The recently published EHRA/ESC consensus statement on a standardised bi-atrial regionalisation provides new opportunities for consistent regional quantitative comparisons within individuals, as well as across patients, modalities and centres. However, although designed to promote clarity and universal applicability, manual annotations based on the 15-segments bi-atrial model remain time-consuming and may be influenced by both operator experience and imaging modality. Purpose To further standardise bi-atrial regionalisation and make it more widely accessible, we developed an open-source software solution that automatically divides any bi-atrial geometry according to the proposed regions. In this study, we validate the accuracy of our automated algorithm by comparing its results to manual expert annotations. Methods The atrial division pipeline DIVAID (DIVision of AtrIal Domains) encompasses three stages: standardisation (mesh resampling and vein and valve clipping), orifice annotation and division. To assess the performance of our algorithm, we processed 70 atrial geometries acquired from multiple modalities (CT, MRI, electroanatomical maps) and centres. Manual annotations were conducted by a blinded expert on the standardised geometries to enable direct comparison with the automatic division results. Regional overlap (Dice score) and mean distance between regional boundaries were used as evaluation metrics. Results All geometries were successfully divided according to the 15-segments bi-atrial model. The average computation time was 24.45s for standardisation, 0.13s for orifice annotation and 2.01s for the division process. Manual division averaged 109.46s, making it over 50 times slower than the automatic approach. Overall, 227 of 279 (81%) veins were clipped correctly at the transition to the atrial body. The remaining vein clips were adjusted manually using an integrated interactive tool. All resulting orifices were correctly annotated across the entire dataset (100%). The regional overlap between manual and automatic division results was 0.98 in the LA and 0.92 in the RA. The mean average distance between regional boundaries of the manual and automatic division was 0.50mm in the LA and 1.84mm in the RA. The largest deviations were observed for the lateral vestibule, which may be attributed to the absence of distinct anatomical landmarks in its vicinity, further highlighting the importance of standardised methods and automated algorithms to achieve consistent and reproducible results. Conclusion Our atrial division pipeline showed good agreement with manual expert annotations, supported by robust preprocessing despite substantial anatomical variability. Automated and standardised large-scale analysis and integration of regional data across patients, imaging modalities, centres and computational models through data-driven approaches may unlock key insights for advancing atrial arrhythmia research and personalised treatment planning.